{"id":"1debae5f-4a9a-46c4-8338-0cfe092b8f4f","arxiv_id":"2509.04875","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Hot-Ham combines Gaunt tensor products with a local-coordinate SO(2) convolution to predict DFT Hamiltonians accurately and efficiently across several material classes.","lead":"This paper presents Hot-Ham, a machine-learning model that predicts quantum mechanical Hamiltonians for materials faster than prior equivariant networks. It combines two tensor-product shortcuts and reports state-of-the-art accuracy on graphene, MoS2, twisted layers, and phosphorus, using fewer parameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single final SO(2)/CGTP layer may not generate all antisymmetric tensor components needed for Hamiltonian readout; no ablation or expressivity proof provided.","rationale":"The reader's weakest assumption identifies the most load-bearing architectural risk: the sufficiency of one final CGTP/SO(2) layer to supply all antisymmetric tensors. This is central because Hot-Ham's entire efficiency gain comes from restricting intermediate layers to symmetric-only GTPs. The paper provides no ablation or formal expressivity check for the final layer. My parity analysis shows that antisymmetric l=0 is not required by the Hamiltonian readout, and l>0 antisymmetric irreps are in principle reachable from symmetric inputs, so the concern is not an immediate refutation. However, the final layer's restricted bilinear form (symmetric features × fixed Y_l filters) could still miss needed components, and the reported accuracy on a few datasets does not rule out systematic errors on structures with stronger antisymmetric contributions. The proposed ablation and rank test would settle this. The reader's CONDITIONAL verdict remains appropriate; no new evidence changes it.","tokens_in":14407,"tokens_out":7100,"duration_ms":71502,"concrete_test":"Ablation: train Hot-Ham with CGTP/SO(2) convolution in the last two layers (adjusting width to keep parameter count within 10%) on monolayer MoS2 and bilayer graphene; if MAE improves by >0.02 meV or the one-layer model shows larger errors on l1≠l2 hopping blocks, the single-layer assumption fails. Additionally, for a representative structure, freeze the trained symmetric features after the last GTP layer and compute the linear map from these features to each required antisymmetric irrep of the readout; if the map is rank-deficient for any non-zero required component, that component is unattainable by construction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Hot-Ham's efficiency claim rests on replacing all intermediate CGTPs with symmetric-only GTPs and introducing antisymmetric tensors only in one final SO(2) convolution layer. The paper states this choice explicitly ('we choose to perform CGTP in the last layer through SO(2) convolution') but offers no ablation or formal expressivity argument. While CGTP of symmetric features with symmetric Y_l filters can produce antisymmetric outputs for l3>0 whenever l1+l2+l3 is odd, the final layer is a restricted bilinear operation on the symmetric features already computed; whether it can generate the full set of antisymmetric irreps required by Eq. (1) for every l1,l2 block is unproven. If some antisymmetric components (e.g., those requiring l1+l2+l3 even but with opposite parity, or particular m couplings) are structurally missing, the model would systematically mispredict those Hamiltonian blocks, and the reported MAEs would only hold for datasets where such components are small. This is the central architectural assumption and is not independently verified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Hot-Ham, an E(3)-equivariant message-passing neural network for predicting DFT Hamiltonians. The architecture combines Gaunt tensor products (GTP), which are efficient but generate only symmetric-parity features, with a single final SO(2)-convolution/CGTP layer intended to introduce the antisymmetric-parity tensors needed by the Hamiltonian readout. The authors report O(L^6)-to-O(L^3) or O(L^2 log^2 L) complexity reductions, benchmark on monolayer graphene, monolayer MoS2, and bilayer graphene, and demonstrate transferability to twisted multilayer MoS2, graphene/h-BN heterostructures, and phosphorus allotropes.","tokens_in":14680,"tokens_out":3252,"duration_ms":33142,"significance":"If the reported accuracy and efficiency hold, Hot-Ham is a meaningful advance: it addresses a real computational bottleneck in equivariant Hamiltonian learning, and the combination of GTP with local-coordinate SO(2) convolution is a plausible design. The generalization tests to unseen twists, heterostructures, and allotropes are valuable and go beyond standard benchmarks. The complexity reduction from O(L^6) to O(L^3) or O(L^2 log^2 L) is well motivated and supported by timing experiments. The paper also reports competitive or superior MAEs with fewer parameters than prior models, which, if reproducible, would make Hot-Ham practically useful for large-scale electronic-structure prediction.","major_comments":[{"comment":"The central architectural assumption is stated but not verified: 'Since only symmetric tensors can be generated by GTP, we need to introduce antisymmetric tensors at least once through CGTP. Given that including antisymmetric tensors in intermediate layers would result in about double computational costs, we choose to perform CGTP in the last layer through SO(2) convolution.' This single final CGTP/SO(2) layer must generate all antisymmetric tensor components required by Eq. (1) for every parity p3 = (-1)^{l1+l2}. The paper gives no expressivity argument and no ablation. If some antisymmetric irreps are structurally missing from this restricted final bilinear map, the model would systematically mispredict those Hamiltonian blocks, and the reported MAEs would be dataset-specific rather than generally accurate. I recommend adding (i) an ablation with antisymmetric tensors in one or more in","section":"Equivariant message passing neural network"},{"comment":"The 'state-of-the-art accuracy' claim is based on single best MAE values compared with the best previously published values from other models. Table 1 reports no error bars, and the Hot-Ham numbers appear to be from a single training run. Since the margins over DeepH-2 and DeePTB are small (e.g., 0.08 vs 0.12 meV for graphene, 0.12 vs 0.14 meV for MoS2), it is unclear whether these differences are statistically significant. Please report the mean and standard deviation over multiple random seeds, and, where feasible, retrain the baselines on the same train/validation/test splits used for Hot-Ham, or at least state the exact split used. Without this, the comparison is suggestive but not conclusive.","section":"Benchmark accuracy / Table 1"}],"minor_comments":[{"comment":"Typo: 'graph neutral networks' should be 'graph neural networks'. Also, 'refered' (Equivariant Hamiltonian section) should be 'referred', and 'traning' (Multilayer twisted MoS2 section) should be 'training'.","section":"Introduction"},{"comment":"The Hermitian enforcement is written as H̃_ij = 1/2 (H_ij + H_ji^⋇). The symbol '⋇' is nonstandard; presumably this is a conjugate transpose or dagger. Please use standard notation, e.g., H_ji^†.","section":"Readout"},{"comment":"The complexity labels in the figure (O(L^2), O(L^2 log^2 L)) are somewhat detached from the plotted curves. Please make clear which curve corresponds to which complexity, or move the labels to the caption/legend.","section":"Fig. 1c"},{"comment":"Reference 6 is incomplete: it lists authors and title but no journal/arXiv identifier. Please complete it. Also check for formatting issues like 'V ASPKIT' in the Applications section.","section":"References"},{"comment":"No statement is provided about code or data availability. Given the empirical nature of the paper, a code release or at least a detailed data-availability statement would substantially aid reproducibility; this is not a correctness issue but would strengthen the contribution.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The core concern is the unverified expressivity of the single final CGTP/SO(2) layer. This is fixable with an ablation or per-block error analysis, so I do not think rejection is warranted. The benchmark comparison would also benefit from error bars and matched training conditions. I believe the paper can be made publishable after these points are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on Hot-Ham. The core idea is simple and attractive: use GTPs (both the 2D-Fourier and sphere-grid variants) for the expensive symmetric tensor products, then compensate for the missing antisymmetric components with a single final SO(2) convolution layer. That's a real engineering contribution, and the efficiency numbers in Fig. 1c show the complexity reduction is not just theoretical. The paper is also honest about the design choice—it explicitly says antisymmetric tensors are introduced only in the last layer to avoid doubling costs. Good for them for stating it plainly.\n\nThe accuracy results are impressive: 0.08, 0.12, and 0.15 meV MAEs on graphene, MoS2, and bilayer graphene, with parameter counts well below most prior models. The generalization tests on twisted MoS2, graphene/h-BN, and phosphorus allotropes are a nice addition and lend credibility to the claim that the model is transferable.\n\nBut there are soft spots. The most important is the antisymmetric tensor assumption. The paper offers no ablation or formal argument that one final CGTP/SO(2) layer can express all the antisymmetric components required by the Wigner-Eckart decomposition. I don't think this is fatal—the empirical success across several very different systems suggests the assumption holds reasonably well in practice—but it is load-bearing and needs to be checked. A simple ablation (e.g., antisymmetric layers at different depths) or a construction for a Hamiltonian block that provably requires a higher-order antisymmetric coupling would settle it. The stress-test note is right to flag this.\n\nSecond, the benchmarks are single-run best MAEs with no error bars, and comparisons are against best-published values from other papers rather than re-trained baselines. This is common in this subfield but still limits the strength of the 'state-of-the-art' claim. Multiple seeds and a standardized training protocol would make the comparison much cleaner.\n\nThird, I don't see any mention of code or data release. For a methods paper in 2024/2025, that's a serious omission. The community would benefit enormously from using the exact architectures.\n\nOverall, this is a coherent, useful paper with a clear contribution. The soft spots are real but not disqualifying. It deserves a serious referee, though not unconditional acceptance.\n\nRecommendation: send it to peer review with a request for ablations on the parity-compensation layer, multiple seeds, and code release. If they address those, it's a solid publication.","headline":"Hot-Ham is a genuinely useful efficiency combination for equivariant Hamiltonian models, with strong benchmark numbers and an honest but untested design assumption that needs an ablation.","tokens_in":15144,"tokens_out":1468,"would_cite":true,"duration_ms":15333,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Hot-Ham predicts DFT Hamiltonians with the best reported mean absolute errors by replacing Clebsch–Gordan tensor products with Gaunt tensor products, plus one final SO(2) layer to restore antisymmetric parity.","keywords":["E(3)-equivariant neural network","machine-learning Hamiltonian","Gaunt tensor product","DFT Hamiltonian prediction","SO(2) convolution","electronic structure","message passing neural network","twisted moiré materials"],"falsifier":"Train Hot-Ham with zero, one, or several final parity-restoring layers on a dataset where the sign of hopping terms under inversion matters (for example, twisted bilayers) and compare Hamiltonian errors; if the one-layer model is not as accurate as the multi-layer model, the central design assumption fails.","tokens_in":14310,"feed_emoji":"⚛️","tokens_out":7414,"duration_ms":70518,"temperature":0.7,"pith_summary":"Hot-Ham is an E(3)-equivariant message-passing network that predicts density-functional-theory Hamiltonians from atomic positions and species. Its central claim is that the expensive Clebsch–Gordan tensor product can be replaced by Gaunt tensor products, computed in Fourier or sphere-grid bases with fast Fourier transforms, while a single final SO(2)-convolution layer restores the antisymmetric tensor components that GTP cannot generate. On standard benchmarks this yields the best reported Hamiltonian mean absolute errors of 0.08, 0.12, and 0.15 meV for monolayer graphene, monolayer MoS2, and bilayer graphene, using fewer than two million parameters. The method also generalizes to twist angles, heterostructure sizes, and crystal allotropes not seen in training, so the practical payoff is DFT-quality electronic structure at a small fraction of the cost.","feed_headline":"Faster tensor math predicts DFT Hamiltonians to 0.08 meV","feed_subtitle":"Gaunt tensor products plus one parity-restoring layer cut cost from L^6 to L^2 log^2 L and beat larger models.","key_machinery":"The central object is the Gaunt tensor product (GTP): instead of directly summing Clebsch–Gordan coefficients, it multiplies the spherical functions associated with two equivariant features in a spectral or grid basis and converts the product back to spherical-harmonic coefficients, enabling FFT acceleration. GTP carries the bulk of the message passing. The complementary mechanism is a local-coordinate SO(2) convolution in the final layer, which is the only place antisymmetric tensors enter; it rotates features so filters become m=0, turning the CGTP into a cheaper 2D operation. Together they enforce E(3) equivariance while keeping the feature set at O(L^3) or O(L^2 log^2 L).","core_discovery":"Hot-Ham establishes that a Hamiltonian-learning equivariant network can get its high-order equivariant features from the Gaunt tensor product rather than full Clebsch–Gordan coupling, provided one CGTP step is kept at the last layer. The Gaunt tensor product expresses the product of two spherical-harmonic features as a pointwise multiplication in a Fourier or sphere-grid basis, reducing the cost of the tensor product from O(L^6) to O(L^3) or O(L^2 log^2 L). Because Gaunt coefficients vanish for odd l1+l2+l3, GTP cannot produce antisymmetric tensors, so Hot-Ham introduces those exactly once, through an SO(2) convolution after local coordinate alignment. The resulting features are read out as","pith_inferences":["The same GTP-plus-one-CGTP design pattern should apply to other equivariant targets that need odd-parity high-order features, such as spin–orbit-coupled or magnetic Hamiltonians, where the single SO(2) layer may need to be revisited.","The strongest unproven spot is the sufficiency of one final antisymmetric layer; an ablation that moves CGTP layers earlier or adds more of them would directly test whether the parity bottleneck is real.","The complexity reduction is asymptotic and assumes the FFT/grid operations dominate; for practical L values, constant-factor overhead and memory layout may determine which GTP variant wins, so the O(L^2 log^2 L) advantage may not always show in wall-clock time.","If the approach generalizes beyond LCAO-basis Hamiltonians, it could extend the same efficiency gain to other quantum-operator learning tasks; the paper only targets the LCAO basis."],"forward_implications":["Hamiltonian prediction for large systems—the paper demonstrates inference on structures with over 10^4 atoms—becomes practical on a single CPU or GPU node.","Models trained on small, untwisted supercells transfer to twisted multilayers, arbitrary moiré mismatches, and allotropes outside the training set, so band structures of hypothetical materials can be screened without new DFT runs.","The parameter count stays under 2M while beating larger models, which lowers training cost and improves data efficiency.","Because the Hamiltonian is predicted directly, band structures, DOS, and derived quantities follow by diagonalization, and the paper notes forces and electron–phonon couplings become accessible via automatic differentiation."],"supporting_citations":[{"why":"Supplies the local-coordinate transformation that reduces Clebsch–Gordan convolutions to SO(2) convolutions, the method Hot-Ham uses for its final antisymmetric layer.","marker":"[20]"},{"why":"Defines the Gaunt tensor product that Hot-Ham adopts for efficient high-order equivariant feature coupling.","marker":"[23]"},{"why":"Provides the sphere-grid GTP variant used as the second tensor-product implementation in the benchmarks.","marker":"[35]"},{"why":"Supplies the DFT Hamiltonian datasets for monolayer graphene, MoS2, and bilayer graphene, and the reference electronic structure used for benchmarks.","marker":"[10]"},{"why":"Establishes the E(3)-equivariant MPNN formulation for Hamiltonians and contributes a baseline model and dataset comparison.","marker":"[11]"},{"why":"Represents the DeepH-2 baseline that also uses SO(2) convolution; Hot-Ham must match or beat its accuracy.","marker":"[13]"},{"why":"Represents the DeePTB baseline with strong previous accuracy; it is the main comparison point in the benchmark table.","marker":"[14]"},{"why":"Defines the equivariant gate used to introduce nonlinearity without breaking E(3) symmetry.","marker":"[36]"}],"fun_headline_variants":["Gaunt tensor product cuts equivariant tensor cost from L^6 to L^2 log L","Hot-Ham: E(3)-equivariant Hamiltonians at O(L^2 log^2 L) cost","Gaunt tensor product enables fast E(3) Hamiltonian learning","Hot-Ham: Accurate DFT Hamiltonians via Gaunt tensor product","E(3)-equivariant Hamiltonians with Gaunt product at lower cost"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The design relies on the assumption that one last layer of the expensive kind of coupling is enough to supply every parity-odd quantity the Hamiltonian needs, and the paper does not test whether more such layers would change the answer.","fun_headline_variants_meta":{"raw":{"variants":["Gaunt tensor product cuts equivariant tensor cost from L^6 to L^2 log L","Hot-Ham: E(3)-equivariant Hamiltonians at O(L^2 log^2 L) cost","Gaunt tensor product enables fast E(3) Hamiltonian learning","Hot-Ham: Accurate DFT Hamiltonians via Gaunt tensor product","E(3)-equivariant Hamiltonians with Gaunt product at lower cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000845,"raw_usage":{"total_tokens":3554,"prompt_tokens":821,"completion_tokens":2733,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2625}},"tokens_in":565,"tokens_out":2733,"duration_ms":18175,"temperature":1.0,"reasoning_tokens":2625,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:48:29.413158+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Hot-Ham with zero, one, or several final parity-restoring layers on a dataset where the sign of hopping terms under inversion matters (for example, twisted bilayers) and compare Hamiltonian errors; if the one-layer model is not as accurate as the multi-layer model, the central design assumption fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DFT Hamiltonian datasets for monolayer graphene, MoS2, and bilayer graphene, and the reference electronic structure used for benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the E(3)-equivariant MPNN formulation for Hamiltonians and contributes a baseline model and dataset comparison."}],"review_version":1}